Agent2Agent (A2A) protocol· Open protocol
Agent2Agent (A2A) Protocol — How AI Agents Talk to Each Other
Quick answer: A2A is a common language for agent-to-agent hand-offs. An agent publishes a small public profile (an Agent Card) saying what it can do and how to reach it; another agent reads the card, sends a message that opens a task, and follows that task until it finishes. The project's own README describes it as an open protocol for communication and interoperability between opaque agentic applications, hosted as an open source project under the Linux Foundation with Google credited as the contributor. Most small businesses will meet it as a question to ask a vendor, not as something to build.
The problem it solves
Suppose your support chatbot is built on one platform and your shipping partner offers a delivery-status agent built on another. Without a shared standard, someone writes a custom integration for that one pair: how to find the agent, how to phrase a request, how to wait for a slow answer, how to sign in. Multiply by every pair of agents and you have a pile of one-off connectors. A2A proposes one agreed set of rules so any compliant agent can work with any other.
The project describes the remote agent as opaque: the client sees what the agent offers, not how it works inside. That is the point. Each side keeps its own model, memory and tools private.
The building blocks
| Piece | What it is (as the A2A documentation describes it) |
|---|---|
| A2A Client | The application or agent that acts for a user and starts the conversation. |
| A2A Server (remote agent) | An agent that exposes an HTTP endpoint implementing the protocol. |
| Agent Card | A JSON document describing the agent's identity, endpoint, capabilities, skills and authentication requirements. |
| Task | A stateful unit of work with a unique ID and a lifecycle, used for long-running jobs and multi-turn collaboration. |
| Message | One turn of communication, from the "user" or the "agent" role, carrying instructions, context or status. |
| Part | The content inside a message or artifact: text, a file, a link or structured data. |
| Artifact | A concrete output the remote agent produces, such as a document or structured result. |
How a request flows
- Discovery. The client finds the Agent Card. The documentation lists three ways: a well-known address on the agent's domain (
/.well-known/agent-card.json, following RFC 8615), a curated registry, or direct configuration such as a config file. - Authentication. The card says what the agent requires. The specification names schemes including API keys, HTTP authentication, OAuth 2.0, OpenID Connect and mutual TLS.
- Send a message. The client sends a message; the server may answer directly or open a task.
- Follow the task. Task states named in the specification include working, completed, failed, canceled and rejected, plus two paused states, input required and authentication required. The page we read did not return its full state list, so check the current specification for the complete set.
- Get progress. For slow work the agent can stream updates (the specification builds streaming on Server-Sent Events) or call a webhook the client registered, if the card says it supports those.
- Collect the artifact. The finished output arrives as one or more artifacts.
The paused states matter for customer-facing bots. "Input required" is how a remote agent says it needs one more detail, such as an order number, before it can continue; your chatbot then has to ask the customer, not guess.
A2A versus MCP
The two get confused because both are open standards for agent systems. The A2A project's own comparison page separates them by direction: MCP is vertical and deepens a single agent by giving it tools and data; A2A is horizontal and connects separate agents, including ones run by other teams or organizations.
| MCP | A2A | |
|---|---|---|
| Connects an agent to | Tools, data sources, functions | Other autonomous agents |
| What is on the other end | Usually narrow, structured and stateless | Something that reasons, plans and holds multi-turn conversations |
| Typical use | Look up an order, query a database | Delegate a whole task to a partner's agent |
The two are meant to be used together. The A2A documentation gives an auto repair shop example: a customer talks to the shop's manager agent over A2A, a mechanic agent uses diagnostic tools over MCP, and the mechanic orders a part from a supplier's agent over A2A. For the tool side, read Model Context Protocol and function calling.
Who backs it, and what that does and does not tell you
The Linux Foundation announced the Agent2Agent project on 23 June 2025, and on 9 April 2026 reported that it had passed 150 supporting organizations. Those are the foundation's own figures. Support from many organizations is not proof that any particular chatbot platform you use implements the protocol, or implements it well; ask each vendor which version and which features it supports.
Limits and risks
- A card can leak. The discovery documentation warns that Agent Cards may expose internal URLs or describe sensitive skills, and recommends authenticated extended cards and access controls for anything private.
- Another agent is another attack surface. Text coming back from a remote agent is untrusted input for your bot; see prompt injection and LLM security.
- Delegation hides responsibility. If a partner's agent gives the customer a wrong answer through your chatbot, the customer blames you. Keep a human handoff path.
- It is a moving standard. The specification lists versions 0.1.0, 0.2.6, 0.3.0 and 1.0.0; code written for an older one may need changes.
FAQ
What is the Agent2Agent (A2A) protocol?
An open standard for how AI agents from different vendors discover each other, send work and receive results. It defines the Agent Card, tasks, messages and the rules for streaming and authentication.
What is an Agent Card?
A JSON document an agent publishes describing who it is, where to reach it, what it can do and how to authenticate. Clients read it to decide whether the agent fits the job.
Is A2A the same as MCP?
No. MCP connects an agent to tools and data; A2A connects an agent to other agents. They are designed to be used together.
Who created A2A?
The project's README credits Google as the contributor, and it now operates as an open source project under the Linux Foundation, licensed under Apache 2.0.
Does my customer-support chatbot need A2A?
Only if it must hand work to an agent run by someone else. A single bot with its own tools needs MCP or plain function calling, not A2A.
Is it safe?
The protocol builds on standard web security, but safety depends on how you configure authentication and how you treat what other agents send back.
Related terms
- Model Context Protocol — the standard for connecting an agent to tools and data.
- Multi-agent system — several agents cooperating on one job.
- AI agent — the kind of software A2A connects.
- Function calling — how a model asks for a tool to be run.
- Human handoff — the exit when no agent should answer.
- LLM security — risks that grow when agents trust each other.
Sources
- A2A Protocol, Specification — a2a-protocol.org/latest/specification, read 1 November 2026. Source of the transports, task states named, streaming, push notification and authentication scheme details, and the version list.
- A2A Protocol, Key concepts — a2a-protocol.org/latest/topics/key-concepts, read 1 November 2026. Source of the actor and object definitions.
- A2A Protocol, Agent discovery — a2a-protocol.org/latest/topics/agent-discovery, read 1 November 2026. Source of the three discovery methods, the well-known path and the security cautions.
- A2A Protocol, A2A and MCP — a2a-protocol.org/latest/topics/a2a-and-mcp, read 1 November 2026. Source of the vertical-versus-horizontal comparison and the repair shop example.
- A2A project repository — github.com/a2aproject/A2A, read 1 November 2026. Source of the project description, Apache 2.0 license and Linux Foundation hosting.
- Linux Foundation, Linux Foundation launches the Agent2Agent Protocol project (23 June 2025) — linuxfoundation.org.
- Linux Foundation, A2A Protocol surpasses 150 organizations… (9 April 2026) — linuxfoundation.org.
- Chatbotscape evaluation methodology. /methodology (continuously updated).